Dear MLE School Participants,
the second day of the MLE Summer School 2022 has a focus on implementing Machine Learning techniques. In the last years a wide range of software components for ML has emerged. In particular libraries such as Tensorflow, PyTorch, Keras, and Scikit-learn become very popular and have become indispensable tools for engineers. While there is yet no dedicated programming language for implementing ML algorithms, some languages offer attractive features for this purpose. One of these languages is Julia with its support for numerical computations and inherent elements for concurrent and parallel programming. But ML is not limited to high performance computing. One of the sessions introduces into the realm of ML on Micro-controllers.
The second day starts with a key note discussing a new trend in ML called federated machine learning. It is about training systems across multiple independent decentralized devices. This approach comes with several benefits such as increased data privacy. Prof. Stefan Schulte is heading the institute of Data Engineering at TUHH. His research is located at the crossroad of data engineering and distributed systems, with a particular focus on federated learning.
Abstract of the Keynote:
Today's Machine Learning (ML) approaches are mostly based on a centralized approach, i.e., data is sent to a centralized entity (very often located in the cloud), where ML training is carried out. However, especially in industrial scenarios, companies are very often not keen on sharing their (raw) data with the cloud, especially if ML training and model generation are provided by an external party (e.g., the vendor of a machine).
Federated Learning (FL) offers an alternative approach, by distributing the model generation across different entities. Thus, learning can be conducted close to the data sources, and only the learned model is shared with other entities. This leads to benefits both with regard to data privacy and communication overhead. In this keynote, we will motivate FL, provide some insights on how to use it, and discuss briefly particular use cases in the Industrial Internet.